Machine Learning and Pattern Recognitionrocha/teaching/2018s1/mo444/classes/2… · Machine...

Machine Learning andPattern Recognition

A High Level Overview

Prof. Anderson Rocha(Main bulk of slides kindly provided by Prof. Sandra Avila)

Institute of Computing (IC/Unicamp)

MC886/MO444

REC Dreasoning for complex data

Machine Learning Technique

Input

Target

Output Target?

Supervised Learning


Input

Output ?Target

Unsupervised Learning

Unsupervised Learning


Input Output

The goal of unsupervised learning is to find patterns in the data, and build new and useful representations of it.

Today’s Agenda

● Clustering

○ k-Means Algorithm

○ Optimization Objective

○ Random Initialization

○ Choosing the Number of Clusters

○ k-Means Variations

○ Evaluation Performance Clustering

Clusteringk-Means Algorithm

Did anyone say pizza?

https://www.youtube.com/watch?v=IpGxLWOIZy4

Did anyone say pizza?

k-Means Clustering

Centroids

Clusters

k-Means: Image Segmentation

K = 10 K = 3 K = 2Original

Credit: Christopher Bishop

k-Means: Image Segmentation

K = 10 K = 3 K = 2Original

Credit: Christopher Bishop

ColorQuantization

k-Means Algorithm

1. Define the k centroids.

2. Find the closest centroid & update cluster assignments.

3. Move the centroids to the center of their clusters.

4. Repeat steps 2 and 3 until the centroid stop moving a lot at each iteration.

k-Means Algorithm (Lloyd's Algorithm)

Input:

➔ K (number of clusters)

➔ Training set


Randomly initialize K cluster centroids ∈ ℝn

repeat {for i = 1 to m

:= index (from 1 to K) of cluster centroid closest tofor k = 1 to K

:= mean of points assigned to cluster k}





:= mean of points assigned to cluster k} Move (Update) centroid step

Cluster assignment step

k-Means Algorithm (Complexity)

● Relatively efficient: O(Kmnt)○ K = #clusters○ m = #vectors (samples)○ n = #dimension of vectors○ t = #iterations○ n ≪ m

Clustering Optimization Objective

k-Means Optimization Objective= index of cluster (from 1 to K) to which example is currently assigned

= cluster centroid k

k-Means Optimization Objective= index of cluster (from 1 to K) to which example is currently assigned

= cluster centroid k

= cluster centroid of cluster to which example has been assigned

Optimization objective:

k-Means Optimization Objective




:= mean of points assigned to cluster k}

Clustering Initialization

Random Initialization

Assign each object to a random cluster &

Computes the initial centroid of each cluster.

Should have K < m.

Randomly pick K trainingexamples.

Set equal to theseK examples.

Random Initialization (Forgy, 1965)

Random Initialization

for i = 1 to 100 {Randomly initialize k-Means.

Run k-means. Get .Compute cost function J.

}

Pick clustering that gave lowest cost J( ).

Can we do better?

● One idea for initializing k-Means is to use a farthest-first traversal on the data set, to pick K points that are far away from each other.

Can we do better?

● One idea for initializing k-Means is to use a farthest-first traversal on the data set, to pick K points that are far away from each other.

● However, this is too sensitive to outliers.

Can we do better?

k-Means++ (Arthur & Vassilvitski, 2007)

● It works similarly to the “farthest” heuristic.

● Choose each point at random, with probability proportional to its squared distance from the centers chosen already.

k-Means++ (Arthur & Vassilvitski, 2007)

● It works similarly to the “farthest” heuristic.

● Choose each point at random, with probability proportional to its squared distance from the centers chosen already.

scikit-learn (default)

ClusteringChoosing the number of clusters

What is the right value of K?

Elbow Method

1 2 3 4 5 67 8 9 10Number of clusters (K)

Cos

t Fun

ctio

n J

Elbow Method

Number of clusters (K)

Cos

t Fun

ctio

n J

1 2 3 4 5 67 8 9 10

Elbow Method


Cos

t Fun

ctio

n J

Elbow Method


Cos

t Fun

ctio

n J

“Elbow”

Elbow Method


Cos

t Fun

ctio

n J

Elbow Method


Cos

t Fun

ctio

n J

Q: You find that cost function J is much higher for k = 5 than for k =3. What can you conclude?

k-Means Variations

Mini-batch k-Means

● Uses mini-batches to reduce the computation time, while still attempting to optimize the same objective function.

● Converges faster than k-Means, but the quality of the results is reduced.

k-Medians Clustering

● Instead of calculating the mean for each cluster to determine its centroid, one instead calculates the median.

● Minimizing error over all clusters with respect to the 1-normdistance metric, as opposed to the square of the 2-normdistance metric (which k-Means does).

k-Medoids Clustering

● Instead of calculating the mean for each cluster to determine its centroid, one instead calculates the medoid.

● Minimizing error over all clusters with respect to the 1-normdistance metric.

● In contrast to the k-Means, k-Medoids chooses data points as centroids.

k-Means (top) vs k-Medoids (bottom)

Credit: https://commons.wikimedia.org/wiki/File:K-means_versus_k-medoids.png

k-Means (left) vs k-Medoids (right)

Credit: https://commons.wikimedia.org/wiki/File:K-means_versus_k-medoids.png

Fuzzy Clustering (Soft Clustering)

● Each data point can belong to more than one cluster.

Hierarchical Clustering

● Agglomerative (“bottom up”): each observation starts in its own cluster, and pairs of clusters are merged as one moves up the hierarchy.

● Divisive (“top down”) :all observations start in one cluster, and splits are performed recursively as one moves down the hierarchy.

DBSCAN Clustering

● Density-Based Spatial Clustering of Applications with Noise

● Given a set of points in some space, it groups together points that are closely packed together (points with many nearby neighbors), marking as outliers points that lie alone in low-density regions.

Clustering Performance Evaluation

Clustering Evaluation

● Evaluating the performance of a clustering algorithm is not as trivial as counting the number of errors or the precision and recall of a supervised classification algorithm.

Clustering Evaluation

● Evaluating the performance of a clustering algorithm is not as trivial as counting the number of errors or the precision and recall of a supervised classification algorithm.

● Adjusted Rand index● Mutual Information based scores● Homogeneity, completeness and V-measure● Silhouette Coefficient

Silhouette Coefficient

● The silhouette value is a measure of how similar a sample is to its own cluster (cohesion) compared to other clusters (separation).

● The silhouette ranges from −1 to +1. ○ High value = the clustering configuration is appropriate.○ Low value = the clustering configuration may have too many or

too few clusters.


● The Silhouette Coefficient is defined for each sample and is composed of two scores:○ a: The mean distance between a sample and all other points

in the same cluster.○ b: The mean distance between a sample and all other points

in the next nearest cluster.


● The Silhouette Coefficient s for a single sample is given as:

● The score is bounded between -1 for incorrect clustering and +1 for highly dense clustering (a ≪ b). Scores around zero indicate overlapping clusters.

http://scikit-learn.org/stable/modules/clustering.html#clustering

https://www.naftaliharris.com/blog/visualizing-k-means-clustering/

References

Machine Learning Books

● Pattern Recognition and Machine Learning, Chap. 9 “Mixture Models and EM”

● Pattern Classification, Chap. 10 “Unsupervised Learning and Clustering”

Machine Learning Courses

● https://www.coursera.org/learn/machine-learning, Week 8

Machine Learning and Pattern Recognitionrocha/teaching/2018s1/mo444/classes/2… · Machine...

Documents

Transcript of Machine Learning and Pattern Recognitionrocha/teaching/2018s1/mo444/classes/2… · Machine...